Pore Segmentation of Permeable Concrete CT Images Based on U2-Net+
A method for image segmentation called U2-Net+ with stacked efficient RSU modules was proposed to ad-dress the problems of current mainstream porous concrete CT image pore segmentation methods.The method introduced more up-sampling nodes and skipped connections by stacking efficient RSU modules in the network,and restored more details of feature maps lost in the down-sampling stage.An additional learnable down-sampling operation was added to enhance the network's ability to capture details in the encoding stage.The original network's depth supervision was simpli-fied to avoid negative impacts of low-level feature maps on fused output feature maps.The single standard binary cross-entropy loss function was replaced with a mixed loss function composed of Focal loss and IoU loss,which improves the network's attention to high-noise pores.Finally,due to the dataset characteristics and network improvements,the num-ber of middle channels in each module of the original network could be further reduced,reducing the network volume.Compared to U2-Net†,U2-Net+ achieves mmIoU,PPrecision and FF1score increased from 94.12% ,88.89% and 93.28% to 94.24% ,91.15% and 94.29% ,while maintaining lightweight and fast performance.The comprehensive indicators of U2-Net+ are superior to those of U-Net,U-Net++,U-Net3+,U2-Net,and U2-Net†.Each evaluation metric has been improved by at least 23.29% compared to mainstream threshold segmentation algorithms,which achieves accurate and fast segmentation of porous concrete CT image pores.
CT image of pervious concreteimage segmentationdeep learningU2-Net
万方数据
